projects/linux·blob v5.11 · refs

463 lines · 11.6 KiB

1// SPDX-License-Identifier: GPL-2.0
2
3#include <linux/context_tracking.h>
4#include <linux/entry-common.h>
5#include <linux/highmem.h>
6#include <linux/livepatch.h>
7#include <linux/audit.h>
8
9#include "common.h"
10
11#define CREATE_TRACE_POINTS
12#include <trace/events/syscalls.h>
13
14/* See comment for enter_from_user_mode() in entry-common.h */
15static __always_inline void __enter_from_user_mode(struct pt_regs *regs)
16{
17 arch_check_user_regs(regs);
18 lockdep_hardirqs_off(CALLER_ADDR0);
19
20 CT_WARN_ON(ct_state() != CONTEXT_USER);
21 user_exit_irqoff();
22
23 instrumentation_begin();
24 trace_hardirqs_off_finish();
25 instrumentation_end();
26}
27
28void noinstr enter_from_user_mode(struct pt_regs *regs)
29{
30 __enter_from_user_mode(regs);
31}
32
33static inline void syscall_enter_audit(struct pt_regs *regs, long syscall)
34{
35 if (unlikely(audit_context())) {
36 unsigned long args[6];
37
38 syscall_get_arguments(current, regs, args);
39 audit_syscall_entry(syscall, args[0], args[1], args[2], args[3]);
40 }
41}
42
43static long syscall_trace_enter(struct pt_regs *regs, long syscall,
44 unsigned long work)
45{
46 long ret = 0;
47
48 /*
49 * Handle Syscall User Dispatch. This must comes first, since
50 * the ABI here can be something that doesn't make sense for
51 * other syscall_work features.
52 */
53 if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
54 if (syscall_user_dispatch(regs))
55 return -1L;
56 }
57
58 /* Handle ptrace */
59 if (work & (SYSCALL_WORK_SYSCALL_TRACE | SYSCALL_WORK_SYSCALL_EMU)) {
60 ret = arch_syscall_enter_tracehook(regs);
61 if (ret || (work & SYSCALL_WORK_SYSCALL_EMU))
62 return -1L;
63 }
64
65 /* Do seccomp after ptrace, to catch any tracer changes. */
66 if (work & SYSCALL_WORK_SECCOMP) {
67 ret = __secure_computing(NULL);
68 if (ret == -1L)
69 return ret;
70 }
71
72 /* Either of the above might have changed the syscall number */
73 syscall = syscall_get_nr(current, regs);
74
75 if (unlikely(work & SYSCALL_WORK_SYSCALL_TRACEPOINT))
76 trace_sys_enter(regs, syscall);
77
78 syscall_enter_audit(regs, syscall);
79
80 return ret ? : syscall;
81}
82
83static __always_inline long
84__syscall_enter_from_user_work(struct pt_regs *regs, long syscall)
85{
86 unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
87
88 if (work & SYSCALL_WORK_ENTER)
89 syscall = syscall_trace_enter(regs, syscall, work);
90
91 return syscall;
92}
93
94long syscall_enter_from_user_mode_work(struct pt_regs *regs, long syscall)
95{
96 return __syscall_enter_from_user_work(regs, syscall);
97}
98
99noinstr long syscall_enter_from_user_mode(struct pt_regs *regs, long syscall)
100{
101 long ret;
102
103 __enter_from_user_mode(regs);
104
105 instrumentation_begin();
106 local_irq_enable();
107 ret = __syscall_enter_from_user_work(regs, syscall);
108 instrumentation_end();
109
110 return ret;
111}
112
113noinstr void syscall_enter_from_user_mode_prepare(struct pt_regs *regs)
114{
115 __enter_from_user_mode(regs);
116 instrumentation_begin();
117 local_irq_enable();
118 instrumentation_end();
119}
120
121/* See comment for exit_to_user_mode() in entry-common.h */
122static __always_inline void __exit_to_user_mode(void)
123{
124 instrumentation_begin();
125 trace_hardirqs_on_prepare();
126 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
127 instrumentation_end();
128
129 user_enter_irqoff();
130 arch_exit_to_user_mode();
131 lockdep_hardirqs_on(CALLER_ADDR0);
132}
133
134void noinstr exit_to_user_mode(void)
135{
136 __exit_to_user_mode();
137}
138
139/* Workaround to allow gradual conversion of architecture code */
140void __weak arch_do_signal_or_restart(struct pt_regs *regs, bool has_signal) { }
141
142static void handle_signal_work(struct pt_regs *regs, unsigned long ti_work)
143{
144 if (ti_work & _TIF_NOTIFY_SIGNAL)
145 tracehook_notify_signal();
146
147 arch_do_signal_or_restart(regs, ti_work & _TIF_SIGPENDING);
148}
149
150static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
151 unsigned long ti_work)
152{
153 /*
154 * Before returning to user space ensure that all pending work
155 * items have been completed.
156 */
157 while (ti_work & EXIT_TO_USER_MODE_WORK) {
158
159 local_irq_enable_exit_to_user(ti_work);
160
161 if (ti_work & _TIF_NEED_RESCHED)
162 schedule();
163
164 if (ti_work & _TIF_UPROBE)
165 uprobe_notify_resume(regs);
166
167 if (ti_work & _TIF_PATCH_PENDING)
168 klp_update_patch_state(current);
169
170 if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
171 handle_signal_work(regs, ti_work);
172
173 if (ti_work & _TIF_NOTIFY_RESUME) {
174 tracehook_notify_resume(regs);
175 rseq_handle_notify_resume(NULL, regs);
176 }
177
178 /* Architecture specific TIF work */
179 arch_exit_to_user_mode_work(regs, ti_work);
180
181 /*
182 * Disable interrupts and reevaluate the work flags as they
183 * might have changed while interrupts and preemption was
184 * enabled above.
185 */
186 local_irq_disable_exit_to_user();
187 ti_work = READ_ONCE(current_thread_info()->flags);
188 }
189
190 /* Return the latest work state for arch_exit_to_user_mode() */
191 return ti_work;
192}
193
194static void exit_to_user_mode_prepare(struct pt_regs *regs)
195{
196 unsigned long ti_work = READ_ONCE(current_thread_info()->flags);
197
198 lockdep_assert_irqs_disabled();
199
200 if (unlikely(ti_work & EXIT_TO_USER_MODE_WORK))
201 ti_work = exit_to_user_mode_loop(regs, ti_work);
202
203 arch_exit_to_user_mode_prepare(regs, ti_work);
204
205 /* Ensure that the address limit is intact and no locks are held */
206 addr_limit_user_check();
207 kmap_assert_nomap();
208 lockdep_assert_irqs_disabled();
209 lockdep_sys_exit();
210}
211
212/*
213 * If SYSCALL_EMU is set, then the only reason to report is when
214 * SINGLESTEP is set (i.e. PTRACE_SYSEMU_SINGLESTEP). This syscall
215 * instruction has been already reported in syscall_enter_from_user_mode().
216 */
217static inline bool report_single_step(unsigned long work)
218{
219 if (work & SYSCALL_WORK_SYSCALL_EMU)
220 return false;
221
222 return work & SYSCALL_WORK_SYSCALL_EXIT_TRAP;
223}
224
225static void syscall_exit_work(struct pt_regs *regs, unsigned long work)
226{
227 bool step;
228
229 /*
230 * If the syscall was rolled back due to syscall user dispatching,
231 * then the tracers below are not invoked for the same reason as
232 * the entry side was not invoked in syscall_trace_enter(): The ABI
233 * of these syscalls is unknown.
234 */
235 if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
236 if (unlikely(current->syscall_dispatch.on_dispatch)) {
237 current->syscall_dispatch.on_dispatch = false;
238 return;
239 }
240 }
241
242 audit_syscall_exit(regs);
243
244 if (work & SYSCALL_WORK_SYSCALL_TRACEPOINT)
245 trace_sys_exit(regs, syscall_get_return_value(current, regs));
246
247 step = report_single_step(work);
248 if (step || work & SYSCALL_WORK_SYSCALL_TRACE)
249 arch_syscall_exit_tracehook(regs, step);
250}
251
252/*
253 * Syscall specific exit to user mode preparation. Runs with interrupts
254 * enabled.
255 */
256static void syscall_exit_to_user_mode_prepare(struct pt_regs *regs)
257{
258 unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
259 unsigned long nr = syscall_get_nr(current, regs);
260
261 CT_WARN_ON(ct_state() != CONTEXT_KERNEL);
262
263 if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
264 if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr))
265 local_irq_enable();
266 }
267
268 rseq_syscall(regs);
269
270 /*
271 * Do one-time syscall specific work. If these work items are
272 * enabled, we want to run them exactly once per syscall exit with
273 * interrupts enabled.
274 */
275 if (unlikely(work & SYSCALL_WORK_EXIT))
276 syscall_exit_work(regs, work);
277}
278
279static __always_inline void __syscall_exit_to_user_mode_work(struct pt_regs *regs)
280{
281 syscall_exit_to_user_mode_prepare(regs);
282 local_irq_disable_exit_to_user();
283 exit_to_user_mode_prepare(regs);
284}
285
286void syscall_exit_to_user_mode_work(struct pt_regs *regs)
287{
288 __syscall_exit_to_user_mode_work(regs);
289}
290
291__visible noinstr void syscall_exit_to_user_mode(struct pt_regs *regs)
292{
293 instrumentation_begin();
294 __syscall_exit_to_user_mode_work(regs);
295 instrumentation_end();
296 __exit_to_user_mode();
297}
298
299noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs)
300{
301 __enter_from_user_mode(regs);
302}
303
304noinstr void irqentry_exit_to_user_mode(struct pt_regs *regs)
305{
306 instrumentation_begin();
307 exit_to_user_mode_prepare(regs);
308 instrumentation_end();
309 __exit_to_user_mode();
310}
311
312noinstr irqentry_state_t irqentry_enter(struct pt_regs *regs)
313{
314 irqentry_state_t ret = {
315 .exit_rcu = false,
316 };
317
318 if (user_mode(regs)) {
319 irqentry_enter_from_user_mode(regs);
320 return ret;
321 }
322
323 /*
324 * If this entry hit the idle task invoke rcu_irq_enter() whether
325 * RCU is watching or not.
326 *
327 * Interrupts can nest when the first interrupt invokes softirq
328 * processing on return which enables interrupts.
329 *
330 * Scheduler ticks in the idle task can mark quiescent state and
331 * terminate a grace period, if and only if the timer interrupt is
332 * not nested into another interrupt.
333 *
334 * Checking for rcu_is_watching() here would prevent the nesting
335 * interrupt to invoke rcu_irq_enter(). If that nested interrupt is
336 * the tick then rcu_flavor_sched_clock_irq() would wrongfully
337 * assume that it is the first interupt and eventually claim
338 * quiescent state and end grace periods prematurely.
339 *
340 * Unconditionally invoke rcu_irq_enter() so RCU state stays
341 * consistent.
342 *
343 * TINY_RCU does not support EQS, so let the compiler eliminate
344 * this part when enabled.
345 */
346 if (!IS_ENABLED(CONFIG_TINY_RCU) && is_idle_task(current)) {
347 /*
348 * If RCU is not watching then the same careful
349 * sequence vs. lockdep and tracing is required
350 * as in irqentry_enter_from_user_mode().
351 */
352 lockdep_hardirqs_off(CALLER_ADDR0);
353 rcu_irq_enter();
354 instrumentation_begin();
355 trace_hardirqs_off_finish();
356 instrumentation_end();
357
358 ret.exit_rcu = true;
359 return ret;
360 }
361
362 /*
363 * If RCU is watching then RCU only wants to check whether it needs
364 * to restart the tick in NOHZ mode. rcu_irq_enter_check_tick()
365 * already contains a warning when RCU is not watching, so no point
366 * in having another one here.
367 */
368 lockdep_hardirqs_off(CALLER_ADDR0);
369 instrumentation_begin();
370 rcu_irq_enter_check_tick();
371 trace_hardirqs_off_finish();
372 instrumentation_end();
373
374 return ret;
375}
376
377void irqentry_exit_cond_resched(void)
378{
379 if (!preempt_count()) {
380 /* Sanity check RCU and thread stack */
381 rcu_irq_exit_check_preempt();
382 if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
383 WARN_ON_ONCE(!on_thread_stack());
384 if (need_resched())
385 preempt_schedule_irq();
386 }
387}
388
389noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
390{
391 lockdep_assert_irqs_disabled();
392
393 /* Check whether this returns to user mode */
394 if (user_mode(regs)) {
395 irqentry_exit_to_user_mode(regs);
396 } else if (!regs_irqs_disabled(regs)) {
397 /*
398 * If RCU was not watching on entry this needs to be done
399 * carefully and needs the same ordering of lockdep/tracing
400 * and RCU as the return to user mode path.
401 */
402 if (state.exit_rcu) {
403 instrumentation_begin();
404 /* Tell the tracer that IRET will enable interrupts */
405 trace_hardirqs_on_prepare();
406 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
407 instrumentation_end();
408 rcu_irq_exit();
409 lockdep_hardirqs_on(CALLER_ADDR0);
410 return;
411 }
412
413 instrumentation_begin();
414 if (IS_ENABLED(CONFIG_PREEMPTION))
415 irqentry_exit_cond_resched();
416 /* Covers both tracing and lockdep */
417 trace_hardirqs_on();
418 instrumentation_end();
419 } else {
420 /*
421 * IRQ flags state is correct already. Just tell RCU if it
422 * was not watching on entry.
423 */
424 if (state.exit_rcu)
425 rcu_irq_exit();
426 }
427}
428
429irqentry_state_t noinstr irqentry_nmi_enter(struct pt_regs *regs)
430{
431 irqentry_state_t irq_state;
432
433 irq_state.lockdep = lockdep_hardirqs_enabled();
434
435 __nmi_enter();
436 lockdep_hardirqs_off(CALLER_ADDR0);
437 lockdep_hardirq_enter();
438 rcu_nmi_enter();
439
440 instrumentation_begin();
441 trace_hardirqs_off_finish();
442 ftrace_nmi_enter();
443 instrumentation_end();
444
445 return irq_state;
446}
447
448void noinstr irqentry_nmi_exit(struct pt_regs *regs, irqentry_state_t irq_state)
449{
450 instrumentation_begin();
451 ftrace_nmi_exit();
452 if (irq_state.lockdep) {
453 trace_hardirqs_on_prepare();
454 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
455 }
456 instrumentation_end();
457
458 rcu_nmi_exit();
459 lockdep_hardirq_exit();
460 if (irq_state.lockdep)
461 lockdep_hardirqs_on(CALLER_ADDR0);
462 __nmi_exit();
463}